Autonomous driving requires detailed and highly accurate maps. However, generating detailed and highly accurate maps can be time consuming and labor and capital intensive. For example, the traditional approach to generating a map of highway overpass locations using global positioning system (GPS)-tagged images begins by training an image-based highway overpass detector. The training process tends to involve the tedious manual annotation of all highway overpass regions within, for example, thousands of images. Thus, before the detector can even be used, a large time, labor, and capital investment is required.